A Bayesian classification of pedestrians in urban areas: The importance of the data preprocessing

We have previously proposed a Bayesian framework to fuse at feature level information from a lidar and video camera in order to classify pedestrians. After studying the influence of each stage of the computation on the system performance, it appears that object segmentation and sensor models are essential for good results. In this paper, we propose some improvements on these steps. Instead of using only the lidar for segmentation, image candidates help to discard the lidar detected objects occluded due to the perspective projection. This leads to a drastic reduction of the number of objects to classify. As a result, the false positive rate decreases and the system is sped up. New sensor models, required in the Bayesian classifier, are also introduced. These distributions come from training processes with databases obtained using our experimental setup. Receiver operating characteristic curves (ROCs) show the ldquooptimalrdquo performances expected while using these distributions in the Bayesian classifier. The evaluation of the proposed object segmentation algorithm on cluttered environments indicates that using only the lidar-based detector increases the number of objects requiring classification by 70%. Experiments, performed on real data in complex urban areas, confirm the effectiveness of the proposed approach.